Intelligent signal analysis methodologies for nuclear detection, identification and attribution

Miltiadis Alamaniotis, Purdue University

Abstract

Detection and identification of special nuclear materials can be fully performed with a radiation detector-spectrometer. Due to several physical and computational limitations, development of fast and accurate radioisotope identifier (RIID) algorithms is essential for automated radioactive source detection and characterization. The challenge is to identify individual isotope signatures embedded in spectral signature aggregation. In addition, background and isotope spectra overlap to further complicate the signal analysis. These concerns are addressed, in this thesis, through a set of intelligent methodologies recognizing signature spectra, background spectrum and, subsequently, identifying radionuclides. Initially, a method for detection and extraction of signature patterns is accomplished by means of fuzzy logic. The fuzzy logic methodology is applied on three types of radiation signal processing applications, where it exhibits high positive detection, low false alarm rate and very short execution time, while outperforming the maximum likelihood fitting approach. In addition, an innovative Pareto optimal multiobjective fitting of gamma ray spectra using evolutionary computing is presented. The methodology exhibits perfect identification while performs better than single objective fitting. Lastly, an innovative kernel based machine learning methodology was developed for estimating natural background spectrum in gamma ray spectra. The novelty of the methodology lies in the fact that it implements a data based approach and does not require any explicit physics modeling. Results show that kernel based method adequately estimates the gamma background, but algorithm's performance exhibits a strong dependence on the selected kernel.

Degree

Ph.D.

Advisors

Tsoukalas, Purdue University.

Subject Area

Nuclear engineering|Artificial intelligence|Computer science

COinS